LaS-Comp: Zero-shot 3D Completion with Latent-Spatial Consistency

Conference: CVPR 2026 arXiv: 2602.18735 Code: https://github.com/wylyan/LaS-Comp Area: Self-Supervised Learning Keywords: 3D shape completion, zero-shot, 3D foundation models, latent-spatial consistency, point cloud completion

TL;DR

This paper proposes LaS-Comp, a zero-shot, category-agnostic 3D shape completion framework. It injects known geometry in the spatial domain via an Explicit Replacement Stage (ERS) and optimizes boundary consistency in the latent space via gradient-based updates in an Implicit Alignment Stage (IAS). The framework bridges the gap between the latent space and spatial domain of pretrained 3D foundation models, achieving state-of-the-art performance across diverse partial observation patterns.

Background & Motivation

3D shape completion is a fundamental problem in computer vision and graphics, aiming to reconstruct complete 3D shapes from partial observations. It has broad applications in robotics, autonomous driving, and AR/VR. An ideal completion method should: (i) robustly handle diverse partial missing patterns (single-view scans, random crops, semantic part removal); (ii) generalize across categories; (iii) not rely on paired data; and (iv) support both text-guided and automatic completion.

Traditional supervised methods depend on paired data and fail to generalize to unseen categories. Recent methods leveraging generative priors (SDS-Complete, ComPC, GenPC) rely on the assumption that the partial input can render at least one complete image — when missing regions are visible from all viewpoints, incomplete renderings cause severe degradation.

The latest 3D foundation models (TRELLIS, Direct3D-S2) adopt a latent-space generation pipeline: shapes are first encoded into a compact latent space via a VAE, and then diffusion/flow-matching models are trained in that latent space. This introduces a unique challenge: even when the complete shape and its partial counterpart share identical geometry in overlapping regions, their latent-space encodings differ significantly. Direct latent-space completion is therefore unreliable.

The core idea of LaS-Comp is to bridge the domain gap between the latent and spatial domains through explicit spatial-domain replacement combined with implicit latent-space alignment, thereby unlocking the completion potential of 3D foundation models.

Method

Overall Architecture

Starting from Gaussian noise, the framework iteratively denoises over multiple steps, guided by the partial input \(\boldsymbol{S}_p\), to progressively recover the complete geometry. At each step \(t \in [0,1]\), two complementary stages are applied: 1. Explicit Replacement Stage (ERS): Explicitly injects geometric information from \(\boldsymbol{S}_p\) in the spatial domain, producing updated latent features \(\boldsymbol{x}_t^*\). 2. Implicit Alignment Stage (IAS): Optimizes \(\boldsymbol{x}_t^*\) via geometric alignment loss gradients, producing spatially aligned features \(\boldsymbol{x}_{t-dt}\).

The complete shape is finally obtained by decoding \(\boldsymbol{S}_c = \mathcal{D}(\boldsymbol{x}_0)\).

Key Designs

1. Explicit Replacement Stage (ERS):

   • Function: Explicitly injects geometric information from the partial input into the latent features of the generation process, ensuring faithful preservation of known regions.
   • Mechanism: Operates via two parallel branches — a Clean Branch and a Noisy Branch.
     • Clean Branch: The generator predicts a noise-free latent \(\hat{\boldsymbol{x}}_{0|t} = \boldsymbol{x}_t - t \cdot \mathcal{G}(\boldsymbol{x}_t, t)\), which is decoded into a complete shape \(\boldsymbol{S}_{0|t}\). A spatial mask replacement is then applied: \(\boldsymbol{S}'_{0|t} = \boldsymbol{S}_p \odot \boldsymbol{M} + \boldsymbol{S}_{0|t} \odot (1-\boldsymbol{M})\), and the result is re-encoded into the latent space as \(\boldsymbol{x}^*_{0|t} = \mathcal{E}(\boldsymbol{S}'_{0|t})\).
     • Noisy Branch + Partial-aware Noise Schedule (PNS): Time-dependent, decreasing perturbations are applied to observed regions (\(\boldsymbol{M}=1\)) to maintain stability, while pure Gaussian noise is applied to missing regions (\(\boldsymbol{M}=0\)) to encourage diverse exploration: \(\boldsymbol{x}^*_{1|t} = \boldsymbol{M} \odot (\sqrt{1-t} \cdot \hat{\boldsymbol{x}}_{1|t} + \sqrt{t} \cdot \boldsymbol{\epsilon}_1) + (1-\boldsymbol{M}) \odot \boldsymbol{\epsilon}_2\)
   • The two branches are combined via flow interpolation: \(\boldsymbol{x}^*_t = (1-t) \cdot \boldsymbol{x}^*_{0|t} + t \cdot \boldsymbol{x}^*_{1|t}\)
   • Design Motivation: Direct latent-space replacement fails due to the domain gap; explicit spatial-domain replacement followed by re-encoding circumvents this issue. PNS assigns different degrees of stochasticity to known and unknown regions.

2. Implicit Alignment Stage (IAS):

   • Function: Corrects discontinuities at the boundary between observed and generated regions that ERS may introduce.
   • Mechanism: Starting from \(\boldsymbol{x}_t^*\), a noise-free latent is predicted and decoded, then a geometric alignment loss is computed: \(\mathcal{L}_{\text{align}} = \text{BCE}(\boldsymbol{S}_{0|t} \odot \boldsymbol{M}, \boldsymbol{S}_p \odot \boldsymbol{M})\). A single-step gradient update is applied to the latent features: \(\boldsymbol{x}^{\text{aligned}}_{0|t} = \hat{\boldsymbol{x}}_{0|t} - \eta \cdot \nabla_{\hat{\boldsymbol{x}}_{0|t}} \mathcal{L}_{\text{align}}\)
   • Design Motivation: While ERS guarantees fidelity, it may introduce artifacts at region boundaries. IAS smooths these inconsistencies via gradient optimization in the latent space. Only a single update step is required, keeping computational overhead minimal.
   • Note: This loss updates only the latent features themselves, not the model parameters.

3. Omni-Comp Benchmark:

   • Function: A new comprehensive evaluation benchmark comprising 30 objects from diverse categories, 3 partial missing patterns (single-view scan, random crop, semantic part removal), and 180 samples in total.
   • Data Sources: 10 Redwood real-world scans + 10 YCB everyday objects + 10 synthetic shapes.
   • Design Motivation: Existing benchmarks are limited in scale (Redwood: only 10 objects), category diversity (KITTI/ScanNet: ≤2 categories), and partial pattern variety (single pattern only).

Loss & Training

LaS-Comp is training-free — it requires no additional training and directly leverages pretrained 3D foundation models (TRELLIS or Direct3D-S2) at inference time. The gradient update learning rate in IAS is \(\eta = 1 \times 10^{-5}\). Completion of each shape takes approximately 20 seconds, more than 3× faster than existing zero-shot methods.

Key Experimental Results

Main Results

Dataset / Metric	Ours (TRELLIS)	Prev. SOTA (ComPC)	Gain
Redwood CD↓/EMD↓	1.42/1.84	1.95/2.59	27.2%/29.0%
Synthetic CD↓/EMD↓	1.11/1.41	1.61/2.09	31.1%/32.5%
ScanNet-Chair UCD↓/UHD↓	0.8/2.0	2.0/5.3	60%/62%
KITTI-Car UCD↓/UHD↓	1.4/4.5	1.1/5.7	—
Omni-Comp Single Scan CD↓	2.21	4.24	47.9%
Omni-Comp Random Crop CD↓	2.60	5.48	52.6%
Omni-Comp Semantic Part CD↓	3.30	6.37	48.2%

Ablation Study

Configuration	Redwood CD↓	Notes
Full LaS-Comp (TRELLIS)	1.42	Best
ERS only (w/o IAS)	1.68	IAS yields ~15% improvement
IAS only (w/o ERS)	2.31	ERS is the core component
w/o PNS (uniform noise schedule)	1.89	PNS is important
Direct3D-S2 backbone	1.64	TRELLIS performs better

Key Findings

• On Omni-Comp, prior methods suffer drastic performance degradation on Random Crop and Semantic Part patterns (due to reliance on the "at least one complete viewpoint" assumption), whereas LaS-Comp remains robust across all patterns.
• The framework is compatible with different 3D foundation models (TRELLIS and Direct3D-S2) and significantly outperforms baselines with both backbones.
• Text-guided completion is supported via the foundation model's CFG mechanism, enabling semantic control over generation results.
• Inference speed is approximately 20 seconds per shape, substantially faster than ComPC (~60s) and SDS-Complete (>5min).

Highlights & Insights

• Training-free: No additional training is required; the method directly leverages pretrained models, making deployment highly practical.
• Addresses a fundamental domain gap: The paper identifies and resolves the previously underappreciated discrepancy between latent encodings of partial versus complete shapes.
• Elegant complementarity of ERS and IAS: Explicit replacement ensures fidelity; implicit alignment ensures smoothness. Both components are indispensable.
• Partial-aware Noise Schedule: Applying distinct noise strategies to known and unknown regions reflects a deep understanding of the asymmetric nature of the completion task.
• Omni-Comp benchmark fills a critical gap in multi-pattern completion evaluation.

Limitations & Future Work

• Performance is bounded by the quality of the pretrained 3D foundation model; weak generative capability for certain categories will limit completion quality.
• Each step requires an encode-decode round trip (ERS), increasing inference overhead.
• IAS performs only a single gradient update step; additional optimization steps could further improve boundary quality at the cost of increased runtime.
• The method has only been validated on rigid objects; articulated and deformable objects remain unexplored.

Related Work & Insights

• RepPaint/FlowDPS inspired the clean-noisy dual-branch design of ERS, transferring concepts from 2D image inpainting to 3D.
• The latent-space generation paradigm of TRELLIS/Direct3D-S2 and similar 3D foundation models provides the foundation for this work.
• The key distinction from ComPC/GenPC is that LaS-Comp does not rely on 2D rendering assumptions and operates directly in the 3D latent space.
• Insight: Latent-spatial consistency may be an equally critical challenge in other latent-space generation tasks, such as 3D editing and style transfer.

Rating

• Novelty: ⭐⭐⭐⭐⭐ First systematic solution to the completion problem for latent-space 3D foundation models; ERS+IAS design is highly original.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Evaluated on 4 existing benchmarks plus the proposed Omni-Comp, with 2 backbones and multiple partial patterns.
• Writing Quality: ⭐⭐⭐⭐ Clear structure, detailed method description, and intuitive figures.
• Value: ⭐⭐⭐⭐⭐ High practical utility as a training-free solution; Omni-Comp benchmark has lasting value for the community.

Related Papers

• [ICCV 2025] CObL: Toward Zero-Shot Ordinal Layering without User Prompting
• [CVPR 2026] An Optimal Transport-driven Approach for Cultivating Latent Space in Online Incremental Learning
• [CVPR 2026] TALO: Pushing 3D Vision Foundation Models Towards Globally Consistent Online Reconstruction
• [CVPR 2026] Breaking the Tuning Barrier: Zero-Hyperparameters Yield Multi-Corner Analysis Via Learned Priors
• [CVPR 2026] Zero-Ablation Overstates Register Content Dependence in DINO Vision Transformers